Abstract
Epstein-Barr virus (EBV) associated gastric cancer, accounting for ~ 9% of all gastric cancers, has unique pathologic, genomic, and clinical features and is linked to a better prognosis. Therefore, we aim to develop and validate a robust deep learning-based classifier to detect EBV positivity for both biopsy and surgical resection specimens in an accurate and reliable manner. We propose a two-stage stain normalization-based robust artificial intelligence classifier for EBV-gastric cancer positivity detection (EBV-TRACER). The results were aggregated to compute three prediction scores including EBV positive cancer-to-tissue ratio, EBV positive cancer-to-tumor ratio, and EBV positive cancer size. To assess the classification performance, we measured the area under the receiver operating curve for detecting EBV status. We also conducted nuclei segmentation and classification to quantitatively analyze and visualize the relationship between the EBV status and lymphocyte features. Our analysis included 2684 gastric specimens with diagnostic clinical reports for gastric cancer collected from January 1, 2011 to December 31, 2023. In the internal validation cohorts, EBV-TRACER yielded AUCs ranging from 0.6596 (95% CI: 0.5359-0.7697) to 0.8414 (95% CI: 0.7267-0.9263) using the three scores. In the external validation cohort, AUCs of 0.7644 (95% CI: 0.6829-0.8399), 0.7652 (95% CI: 0.6850-0.8398), and 0.7221 (95% CI: 0.6134-0.8178) were obtained for the three scores, respectively. Overall, EBV-TRACER significantly outperforms models without stain normalization and those using conventional stain normalization. It could serve as a promising tool for improving the efficiency and accuracy of decision making in gastric cancer diagnosis, supporting more effective treatment planning and strategy development.